The recent issue of Biology Bulletin (2024, Vol. 51, No. 5) published an article entitled “Automated Identification and Counting of Saigas (Saiga tatarica) by Using Deep Convolutional Neural Networks in High-Resolution Satellite Images” which presents the results of work on recognizing saigas in images of the Earth from space using artificial intelligence. For this work, the entire data set (high-detail images) for the North-West Caspian region, stored at the IEE RAS throughout all the years of work on this topic, was used. High-resolution images (0.5 m per pixel) of past years were accumulated for the period from 2012 to 2022, they captured the same territory in the same season of the year, namely the protected areas of the North-West Caspian region. The Institute has been successfully cooperating with the Stepnoy Nature Reserve in the Astrakhan Region for many years. Joint activities of specialists allow for prompt and accurate implementation of work on verification of data found in space images - on the ground, in the steppe. Thus, employees of the Reserve can at any time confirm or refute the assumptions of the decoders of space images at the Institute on their territory. So the data obtained during the analysis of these images in previous years are highly reliable. Previously, specialists of the Institute developed a method for counting saigas from space - in 2015 it was published in the article "On the Possibility to Identify the Saiga Antelope (Saiga tatarica) on Veryhigh Resolution Satellite Images".
Space images have been actively used to assess the state of animal habitats for the last 18-20 years, which is in some way related to the assessment of the number of endangered species (however, this is an indirect connection). The first works with space images with a resolution that allows one to distinguish the animal itself on it (no coarser than 0.5-1 m per pixel) were published in 2014-2015 (respectively, the time of the appearance of satellites with the appropriate equipment and the possibility of purchasing such images for a wide contingent of users). Such works are most optimal for animals of open spaces that are large in size: elephants in the savannah, elephant seals and walruses in open haulouts, polar bears in the Arctic ice desert.
The first work by employees of the IEE RAS on research devoted to distinguishing saigas on space images was also published in 2015, which has already been mentioned, and they began to develop this direction in 2012.
“At that time, it was not easy, since saigas are graceful antelopes, and much smaller than elephants or walruses in size. Since the ideas of using Artificial Intelligence are currently being intensively developed, the previously developed method also experienced an urgent need to attract modern capabilities; after all, the areas in which saigas need to be detected and counted are huge. In 2023, we began work on selecting neural network algorithms appropriate for the tasks and training them on the existing data pool - that is, on space images that we had previously used to count saigas in the steppe manually, which required certain skills and a lot of time,” says Anna Yachmennikova, PhD in Biology, Senior Researcher at the IEE RAS.
Scientists from the Institute of Ecology and Evolution of the Russian Academy of Sciences work together with programmers and engineers from Es-Paz, with whom they also actively collaborate on issues of satellite telemetry and tracking animals using transmitters. The "Saiga Search Network" was developed thanks to the collaboration of specialists from different fields. The published article, the Russian title of which is translated as "Automated detection and counting of saigas (Saiga tatarica) on super-detailed satellite images using deep convolutional neural networks", describes in detail the methodology of work on writing an algorithm by programmers and verification of data and results at all stages of work by zoologists. Automated counting of animals was done in two stages. At both stages, artificial intelligence technologies were used - deep convolutional neural networks (DCNN), but of different architectures, were used for efficient image processing: the work of one network was aimed at identifying clusters of animals, and the other - at detecting each saiga individually. The neural network that detects clusters was developed based on the standard ResNet-50 architecture, designed to solve image classification problems, and then, using a neural network based on another YOLOv7 architecture (one of the most modern), the problems of detecting objects were solved, and so the second network was developed.
This is our first experience of using AI to count animals in space images. Basically, all such work comes down to working with aerial photography, but this is a completely different level. Previously, saigas were counted using automobiles or aircrafts - small aircrafts. Such counting now has an extremely negative impact on saigas: cars disturb the cover of vulnerable steppe ecosystems of the North-West Caspian region, and aircrafts cause panic flight among the saigas themselves (they may even die from exhaustion). In addition, such methods of counting are not able to cover large areas at one time. Artificial intelligence allows this meticulous and voluminous work to be done much faster, and the results can be used not only for statistics, but also for making operational decisions (if such a need arises).
Images with a resolution of half a meter per pixel are used in this process. Throughout our history, we have worked with satellite systems of Israel, France, and China. Since 2012, IEE RAS has been making an annual request for satellite monitoring of the state of the saiga population in the Northwestern Caspian region. Unfortunately, the request for funding and purchasing satellite images was not supported every year, and images of such high resolution are expensive. If scientists have confirmation of funding for this work, they contact the company purchasing space images in advance and place a special order through it for the company that owns the satellite of interest to scientists. Guided by the terms of the special order, the satellite, at a predetermined time of the year, under cloudless weather, captures the area of interest to the researchers until it receives images that meet the requirements, or until the period of the year during which it is important to conduct such capture ends. The period in which the steppe is photographed corresponds to the saiga rutting season - November-December: this is the time when the males herd the females into harems and the animals stay in large, sedentary groups, and it is relatively easy to count them from above.
The saiga is a species that is prone to sharp population jumps if the combination of factors influencing it is favorable. However, one should not delude oneself with this, if the conditions are unfavorable, saigas quickly and easily die out en masse, in huge numbers. It is necessary to carefully monitor populations and keep these biological characteristics in mind, making timely and competent decisions on working with this species.
Earlier, employees of the IEE RAS conducted work on modeling the fate of this population in the event of a confluence of certain natural circumstances. Under exceptionally favorable conditions for the saiga, growth is described in accordance with the exponent. That is, in the initial period there is a very slow increase in the population, and after the inflection point - very fast and intensive. However, according to the model's forecasts, by 2015 there should have been at least 50,000 animals in the North-West Caspian region. So the fact that in 2022 there were about 25 thousand of them suggests that the negative factors that affected the saiga at the beginning of the reference point (from the moment the population entered a state of extremely low numbers) were very powerful and practically stopped the natural processes of restoration of this population. It is very important that this work be continued, because natural processes are continuous. Monitoring of the population must be constant, otherwise fundamentally important events in the fate of the species can be missed. The more scientists know, the better they can confirm the existing situation with objective data obtained from reality - the less risk there is of making wrong decisions that can lead to the loss of this unique relict species in the territory of the North-West Caspian region in Russia.
The work was published in the journal Biology Bulletin, volume 51, pages 1407–1421, (2024), V. V. Rozhnov, A. L. Salman, A. A. Yachmennikova, A. A. Lushchekina & P. A. Salman.
АвPhoto by: Andrey Giljov
Related materials:
All news from Astrakhan: "Russian scientists use new digital technologies in the Stepnoy nature reserve"
Nature Management and Environmental Protection Service of the Astrakhan Region: "Artificial intelligence helps in counting Astrakhan saigas"